lnet/klnds/socklnd/socklnd_modparams.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2012, Intel Corporation.
   5  *
   6  *   Author: Eric Barton <eric@bartonsoftware.com>
   7  *
   8  *   Portals is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Portals is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Portals; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include "socklnd.h"
  23
  24 #include <linux/kvm_host.h>
  25 #if defined(__x86_64__) || defined(__i386__)
  26 #include <asm/hypervisor.h>
  27 #endif
  28
  29 #define CURRENT_LND_VERSION 1
  30
  31 static int sock_timeout;
  32 module_param(sock_timeout, int, 0644);
  33 MODULE_PARM_DESC(sock_timeout, "dead socket timeout (seconds)");
  34
  35 static int credits = DEFAULT_CREDITS;
  36 module_param(credits, int, 0444);
  37 MODULE_PARM_DESC(credits, "# concurrent sends");
  38
  39 static int peer_credits = DEFAULT_PEER_CREDITS;
  40 module_param(peer_credits, int, 0444);
  41 MODULE_PARM_DESC(peer_credits, "# concurrent sends to 1 peer");
  42
  43 static int peer_buffer_credits;
  44 module_param(peer_buffer_credits, int, 0444);
  45 MODULE_PARM_DESC(peer_buffer_credits, "# per-peer router buffer credits");
  46
  47 static int peer_timeout = DEFAULT_PEER_TIMEOUT;
  48 module_param(peer_timeout, int, 0444);
  49 MODULE_PARM_DESC(peer_timeout, "Seconds without aliveness news to declare peer dead (<=0 to disable)");
  50
  51 /* Number of daemons in each thread pool which is percpt,
  52  * we will estimate reasonable value based on CPUs if it's not set. */
  53 static unsigned int nscheds;
  54 module_param(nscheds, int, 0444);
  55 MODULE_PARM_DESC(nscheds, "# scheduler daemons in each pool while starting");
  56
  57 static int nconnds = 4;
  58 module_param(nconnds, int, 0444);
  59 MODULE_PARM_DESC(nconnds, "# connection daemons while starting");
  60
  61 static int nconnds_max = 64;
  62 module_param(nconnds_max, int, 0444);
  63 MODULE_PARM_DESC(nconnds_max, "max # connection daemons");
  64
  65 static int min_reconnectms = 1000;
  66 module_param(min_reconnectms, int, 0644);
  67 MODULE_PARM_DESC(min_reconnectms, "min connection retry interval (mS)");
  68
  69 static int max_reconnectms = 60000;
  70 module_param(max_reconnectms, int, 0644);
  71 MODULE_PARM_DESC(max_reconnectms, "max connection retry interval (mS)");
  72
  73 static int eager_ack;
  74 module_param(eager_ack, int, 0644);
  75 MODULE_PARM_DESC(eager_ack, "send tcp ack packets eagerly");
  76
  77 static int typed_conns = 1;
  78 module_param(typed_conns, int, 0444);
  79 MODULE_PARM_DESC(typed_conns, "use different sockets for bulk");
  80
  81 static int min_bulk = (1<<10);
  82 module_param(min_bulk, int, 0644);
  83 MODULE_PARM_DESC(min_bulk, "smallest 'large' message");
  84
  85 # define DEFAULT_BUFFER_SIZE 0
  86 static int tx_buffer_size = DEFAULT_BUFFER_SIZE;
  87 module_param(tx_buffer_size, int, 0644);
  88 MODULE_PARM_DESC(tx_buffer_size, "socket tx buffer size (0 for system default)");
  89
  90 static int rx_buffer_size = DEFAULT_BUFFER_SIZE;
  91 module_param(rx_buffer_size, int, 0644);
  92 MODULE_PARM_DESC(rx_buffer_size, "socket rx buffer size (0 for system default)");
  93
  94 static int nagle = 0;
  95 module_param(nagle, int, 0644);
  96 MODULE_PARM_DESC(nagle, "enable NAGLE?");
  97
  98 static int round_robin = 1;
  99 module_param(round_robin, int, 0644);
 100 MODULE_PARM_DESC(round_robin, "Round robin for multiple interfaces");
 101
 102 static int keepalive = 30;
 103 module_param(keepalive, int, 0644);
 104 MODULE_PARM_DESC(keepalive, "# seconds before send keepalive");
 105
 106 static int keepalive_idle = 30;
 107 module_param(keepalive_idle, int, 0644);
 108 MODULE_PARM_DESC(keepalive_idle, "# idle seconds before probe");
 109
 110 #define DEFAULT_KEEPALIVE_COUNT  5
 111 static int keepalive_count = DEFAULT_KEEPALIVE_COUNT;
 112 module_param(keepalive_count, int, 0644);
 113 MODULE_PARM_DESC(keepalive_count, "# missed probes == dead");
 114
 115 static int keepalive_intvl = 5;
 116 module_param(keepalive_intvl, int, 0644);
 117 MODULE_PARM_DESC(keepalive_intvl, "seconds between probes");
 118
 119 static int enable_csum = 0;
 120 module_param(enable_csum, int, 0644);
 121 MODULE_PARM_DESC(enable_csum, "enable check sum");
 122
 123 static int inject_csum_error = 0;
 124 module_param(inject_csum_error, int, 0644);
 125 MODULE_PARM_DESC(inject_csum_error, "set non-zero to inject a checksum error");
 126
 127 static int enable_irq_affinity = 0;
 128 module_param(enable_irq_affinity, int, 0644);
 129 MODULE_PARM_DESC(enable_irq_affinity, "enable IRQ affinity");
 130
 131 static int nonblk_zcack = 1;
 132 module_param(nonblk_zcack, int, 0644);
 133 MODULE_PARM_DESC(nonblk_zcack, "always send ZC-ACK on non-blocking connection");
 134
 135 static unsigned int zc_min_payload = (16 << 10);
 136 module_param(zc_min_payload, int, 0644);
 137 MODULE_PARM_DESC(zc_min_payload, "minimum payload size to zero copy");
 138
 139 static unsigned int zc_recv = 0;
 140 module_param(zc_recv, int, 0644);
 141 MODULE_PARM_DESC(zc_recv, "enable ZC recv for Chelsio driver");
 142
 143 static unsigned int zc_recv_min_nfrags = 16;
 144 module_param(zc_recv_min_nfrags, int, 0644);
 145 MODULE_PARM_DESC(zc_recv_min_nfrags, "minimum # of fragments to enable ZC recv");
 146
 147 static unsigned int conns_per_peer = DEFAULT_CONNS_PER_PEER;
 148 module_param(conns_per_peer, uint, 0644);
 149 MODULE_PARM_DESC(conns_per_peer, "number of connections per peer");
 150
 151 #ifdef SOCKNAL_BACKOFF
 152 static int backoff_init = 3;
 153 module_param(backoff_init, int, 0644);
 154 MODULE_PARM_DESC(backoff_init, "seconds for initial tcp backoff");
 155
 156 static int backoff_max = 3;
 157 module_param(backoff_max, int, 0644);
 158 MODULE_PARM_DESC(backoff_max, "seconds for maximum tcp backoff");
 159 #endif
 160
 161 #if SOCKNAL_VERSION_DEBUG
 162 static int protocol = 3;
 163 module_param(protocol, int, 0644);
 164 MODULE_PARM_DESC(protocol, "protocol version");
 165 #endif
 166
 167 static inline bool is_native_host(void)
 168 {
 169 #ifdef HAVE_HYPERVISOR_IS_TYPE
 170         return hypervisor_is_type(X86_HYPER_NATIVE);
 171 #elif defined(__x86_64__) || defined(__i386__)
 172         return x86_hyper == NULL;
 173 #else
 174         return true;
 175 #endif
 176 }
 177
 178 struct ksock_tunables ksocknal_tunables;
 179 static struct lnet_ioctl_config_socklnd_tunables default_tunables;
 180
 181 int ksocknal_tunables_init(void)
 182 {
 183         default_tunables.lnd_version = CURRENT_LND_VERSION;
 184         default_tunables.lnd_conns_per_peer = conns_per_peer;
 185
 186         /* initialize ksocknal_tunables structure */
 187         ksocknal_tunables.ksnd_timeout            = &sock_timeout;
 188         ksocknal_tunables.ksnd_nscheds            = &nscheds;
 189         ksocknal_tunables.ksnd_nconnds            = &nconnds;
 190         ksocknal_tunables.ksnd_nconnds_max        = &nconnds_max;
 191         ksocknal_tunables.ksnd_min_reconnectms    = &min_reconnectms;
 192         ksocknal_tunables.ksnd_max_reconnectms    = &max_reconnectms;
 193         ksocknal_tunables.ksnd_eager_ack          = &eager_ack;
 194         ksocknal_tunables.ksnd_typed_conns        = &typed_conns;
 195         ksocknal_tunables.ksnd_min_bulk           = &min_bulk;
 196         ksocknal_tunables.ksnd_tx_buffer_size     = &tx_buffer_size;
 197         ksocknal_tunables.ksnd_rx_buffer_size     = &rx_buffer_size;
 198         ksocknal_tunables.ksnd_nagle              = &nagle;
 199         ksocknal_tunables.ksnd_round_robin        = &round_robin;
 200         ksocknal_tunables.ksnd_keepalive          = &keepalive;
 201         ksocknal_tunables.ksnd_keepalive_idle     = &keepalive_idle;
 202         ksocknal_tunables.ksnd_keepalive_count    = &keepalive_count;
 203         ksocknal_tunables.ksnd_keepalive_intvl    = &keepalive_intvl;
 204         ksocknal_tunables.ksnd_credits            = &credits;
 205         ksocknal_tunables.ksnd_peertxcredits      = &peer_credits;
 206         ksocknal_tunables.ksnd_peerrtrcredits     = &peer_buffer_credits;
 207         ksocknal_tunables.ksnd_peertimeout        = &peer_timeout;
 208         ksocknal_tunables.ksnd_enable_csum        = &enable_csum;
 209         ksocknal_tunables.ksnd_inject_csum_error  = &inject_csum_error;
 210         ksocknal_tunables.ksnd_nonblk_zcack       = &nonblk_zcack;
 211         ksocknal_tunables.ksnd_zc_min_payload     = &zc_min_payload;
 212         ksocknal_tunables.ksnd_zc_recv            = &zc_recv;
 213         ksocknal_tunables.ksnd_zc_recv_min_nfrags = &zc_recv_min_nfrags;
 214         if (conns_per_peer > ((1 << SOCKNAL_CONN_COUNT_MAX_BITS)-1)) {
 215                 CWARN("socklnd conns_per_peer is capped at %u.\n",
 216                       (1 << SOCKNAL_CONN_COUNT_MAX_BITS)-1);
 217         }
 218         ksocknal_tunables.ksnd_conns_per_peer     = &conns_per_peer;
 219
 220         if (enable_irq_affinity) {
 221                 CWARN("irq_affinity is removed from socklnd because modern "
 222                       "computer always has fast CPUs and more cores than "
 223                       "# NICs, although you still can set irq_affinity by "
 224                       "another way, please check manual for details.\n");
 225         }
 226         ksocknal_tunables.ksnd_irq_affinity       = &enable_irq_affinity;
 227
 228 #ifdef SOCKNAL_BACKOFF
 229         ksocknal_tunables.ksnd_backoff_init       = &backoff_init;
 230         ksocknal_tunables.ksnd_backoff_max        = &backoff_max;
 231 #endif
 232
 233 #if SOCKNAL_VERSION_DEBUG
 234         ksocknal_tunables.ksnd_protocol           = &protocol;
 235 #endif
 236
 237         if (*ksocknal_tunables.ksnd_zc_min_payload < (2 << 10))
 238                 *ksocknal_tunables.ksnd_zc_min_payload = (2 << 10);
 239
 240         /* When on a hypervisor set the minimum zero copy size
 241          * above the maximum payload size
 242          */
 243         if (!is_native_host())
 244                 *ksocknal_tunables.ksnd_zc_min_payload = (16 << 20) + 1;
 245
 246         return 0;
 247 }
 248
 249 void ksocknal_tunables_setup(struct lnet_ni *ni)
 250 {
 251         struct lnet_ioctl_config_socklnd_tunables *tunables;
 252         struct lnet_ioctl_config_lnd_cmn_tunables *net_tunables;
 253
 254         /* If no tunables specified, setup default tunables */
 255         if (!ni->ni_lnd_tunables_set)
 256                 memcpy(&ni->ni_lnd_tunables.lnd_tun_u.lnd_sock,
 257                        &default_tunables, sizeof(*tunables));
 258
 259         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
 260
 261         /* Current API version */
 262         tunables->lnd_version = CURRENT_LND_VERSION;
 263
 264         net_tunables = &ni->ni_net->net_tunables;
 265
 266         if (net_tunables->lct_peer_timeout == -1)
 267                 net_tunables->lct_peer_timeout =
 268                         *ksocknal_tunables.ksnd_peertimeout;
 269
 270         if (net_tunables->lct_max_tx_credits == -1)
 271                 net_tunables->lct_max_tx_credits =
 272                         *ksocknal_tunables.ksnd_credits;
 273
 274         if (net_tunables->lct_peer_tx_credits == -1)
 275                 net_tunables->lct_peer_tx_credits =
 276                         *ksocknal_tunables.ksnd_peertxcredits;
 277
 278         if (net_tunables->lct_peer_tx_credits >
 279             net_tunables->lct_max_tx_credits)
 280                 net_tunables->lct_peer_tx_credits =
 281                         net_tunables->lct_max_tx_credits;
 282
 283         if (net_tunables->lct_peer_rtr_credits == -1)
 284                 net_tunables->lct_peer_rtr_credits =
 285                         *ksocknal_tunables.ksnd_peerrtrcredits;
 286
 287         if (!tunables->lnd_conns_per_peer)
 288                 tunables->lnd_conns_per_peer = (conns_per_peer) ?
 289                         conns_per_peer : DEFAULT_CONNS_PER_PEER;
 290 }